The Research Of Seismic Inversion Methods Based On Deep Convolutional Neural Network

Seismic inversion is a branch of geophysical inversion,aiming at providing reliable information for oil and gas exploration by obtaining elastic parameters and petrophysical parameters from scant well logging information and seismic records.Nevertheless,the conventional inversion process often faces problems including ill-conditioned matrix,overdetermined equations,slow algorithm convergence,poor stability et cetera.And these problems have bad effects on the accuracy of inversion results.How to improve the accuracy of inversion results and fit the relationship between reservoir parameters and seismic recordings have become the current research focus.DNN(Deep Neural Network)realizes nonlinear mapping between inputs and outputs through the combination of complex network structures and activation functions.It can avoid the process of solving complex relation equation and realize the mapping from seismic recordings to reservoir parameters directly,so that we can get accurate reservoir parameters.In this paper,we mainly focus on the seismic inversion method based on deep learning.We propose two network structures,deep CNN(Convolutional Neural Network)applied for post-stack seismic inversion for P-wave impedance and deep CNN-LSTM(Long Short-Term Memory)applied for azimuthal pre-stack seismic inversion for fracture parameters.Two network both take the convolution layer in CNN as the main structure,using the ability of feature extraction of CNN to extract the feature of seismic recordings.And we propose improved loss functions for two networks separately according to seismic forward modeling process.In post-stack seismic inversion,we build shallow CNN and deep CNN network.Good network parameters are chosen through model tests.Then we apply the best network to real work area data and get good results.In pre-stack seismic inversion,we build deep CNN,deep CNN-LSTM and improve the loss function according to the forward modeling of azimuthal pre-stack seismic.Then we choose the best network for real work area data based on model tests.It turns out that deep CNN works better than shallow CNN and deep CNN-LSTM is more stable than single deep CNN structure.Besides,improved loss functions significantly improve the accuracy of predicted results.